Abstract:
Automated detection and segmentation of fruit-surface defects are essential for improving
product quality, reducing dependence on manual inspection, and enabling scalable agricultural
automation. Fruit defects vary widely in shape, texture, size, and severity, limiting the
generalization of deep learning models. This study evaluates semantic segmentation of
orange surface defects using a U-Net encoder–decoder with a ResNet34 backbone, trained
on the Orange Navel dataset comprising 1,448 images annotated for three defect categories:
sunburn, ulcer, and wind scarring. The model achieves strong quantitative performance:
mean Pixel Accuracy (mPA) of 90.57%, mean Intersection over Union (mIoU) of 84.68%,
and Dice coefficient of 91.37%. Class-wise IoU highlights near-perfect background segmentation
(99.13%) and robust performance on sunburn (85.67%), while ulcer (83.61%) and wind scarring
(70.29%) show reduced fidelity-likely due to subtler texture cues, lower contrast, and more
irregular lesion boundaries. These results position U-Net with a ResNet34 encoder as a
reliable baseline for fruit-defect segmentation. To strengthen generalization under limited
annotations, a semi-supervised learning strategy based on the FixMatch paradigm is employed.
In this approach, pseudo labels are generated for unlabeled samples, and only high-confidence
predictions are used as supervisory signals. Training is guided by enforcing consistency
between the model’s predictions and these pseudo labels, allowing effective use of both labeled
and unlabeled data and improving segmentation performance across all defect categories.
